Abstract

The amount of training data is one of the critical factors affecting the performance of handwritten numeral recognition system. One way to increase the training data size is adding synthesized data. In this study, synthetic data generation using morphing transformation with convex evolution is investigated. This technique uses a pair of original samples as the source and the target, and generates the synthetic samples by evolving the source towards the target. We aim to balance the data distribution. Normally, the training data has poor distribution due to the data redundancy and sparseness caused by frequent and rare samples, respectively. In terms of data clusters, some clusters are small, some are large and filling the gap between these clusters with synthetic data should smooth the clusters. Using the Support Vector Machines method, the rare samples, also called support vectors, are determined. Then, the number of rare samples is increased using morphing transformation. Using this technique a recognition rate of 99.19% has been achieved, while the initial performance without morphing was 99.07%. Morphing transformation generated more representative synthetic samples, which cannot be obtained by the other data synthesis methods such as affine and elastic distortions.